DeepNeXt: a lightweight polyp segmentation algorithm based on multi-scale attention.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Accurate polyp segmentation in colonoscopy is vital for early colon cancer detection and treatment.
• Current deep learning models for polyp segmentation are often too large and lack sufficient accuracy for clinical use.
• There is a clinical need for lightweight, efficient models for integration into medical devices.

Purpose of the Study:

• To develop a lightweight deep neural network for fast and accurate automatic polyp segmentation.
• To create a model suitable for embedding in clinical devices for practical applications.
• To provide technical support for rapid and precise polyp segmentation in clinical settings.

Main Methods:

• Introduced DeepNeXt, a polyp segmentation model utilizing multi-scale attention mechanisms.
• Employed a multi-stage, lightweight convolutional encoder for efficient feature extraction.
• Implemented a multi-stage feature fusion structure to prevent information loss during encoding.
• Utilized multi-scale attentional feature encoding with deep strip convolutions for diverse feature extraction.

Main Results:

• DeepNeXt demonstrated superior performance compared to mainstream networks (U-net, U-net++, TransUnet, SwinUnet, TGANet) on Kvasir-SEG and CVC-ClinicDB datasets.
• Achieved significantly low computational costs: 3.04 G FLOPs and 1.51 M parameters.
• Obtained high segmentation accuracy with mIOU of 83.91% (Kvasir-SEG) and 87.37% (CVC-ClinicDB).
• Exhibited excellent Dice and Recall metrics, balancing efficiency, compactness, and accuracy.

Conclusions:

• Proposed DeepNeXt, a novel lightweight network with multi-scale attention for polyp segmentation.
• The model is specifically designed for computationally limited medical devices.
• DeepNeXt offers strong support for accurate and efficient polyp segmentation in clinical applications.